PSA: Anyone with a link can view your Granola notes by default
If you use the AI-powered note-taking app Granola, you might want to double-check your privacy settings. Though Granola says your notes are "private by default," it makes them viewable to anyone with a link, and also uses them for internal AI training unless you opt out. Granola describes itself as an "AI notepad for people in back-to-back meetings." It integrates with your calendar to capture audio from your meetings, and then uses AI to generate a bulleted list of what you've heard, which it calls a "note." You can edit the AI-generated notes, invite other collaborators to view them, and use Granola's AI assistant to ask questions about y … Read the full story at The Verge.
China’s Five-Year Plan details the targets for AI deployment
China has approved its 15th Five-Year Plan [PDF] setting out the country’s economic, education, social, and industrial priorities through to 2030. AI is grouped alongside quantum computing, biotechnology, and energy as paths that are to be pursued as part of the country’s strategic science policy. The document calls for more work in developing high-performance AI chips and the software to support them in this context. There’s also a commitment to academic and industry research on new model architectures and the core algorithms underpinning them. In the section of the Five-Year Plan dedicated to digital infrastructure, the use of AI falls into three components: computing power, AI models, and the organisation and dissemination of data across China.
Welcome Gemma 4: Frontier multimodal intelligence on device
Table of Contents Multimodal Capabilities Deploy Anywhere Fine-tuning & Demos Fine-tuning with Unsloth Studio Try Gemma 4 Acknowledgements What is new with Gemma 4? Gemma 4 comes in four sizes, all base and instruction fine-tuned: Overview of Capabilities and Architecture Gemma 4 leverages several architecture components used in previous Gemma versions and other open models, and leaves out complex or inconclusive features such as Altup. These are the main architecture characteristics in Gemma 4: Per-Layer Embeddings (PLE) One of the most distinctive features in smaller Gemma 4 models is Per-Layer Embeddings (PLE), which was introduced previously in Gemma-3n. For multimodal inputs (images, audio, video), PLE is computed before soft tokens are merged into the embedding sequence — since PLE relies on token IDs that are lost once multimodal features replace the placeholders.
Holo3: Breaking the Computer Use Frontier
With a score of 78.85% on the OSWorld-Verified benchmark, Holo3-122B-A10B establishes a new state of the art for the industry on the leading desktop computer use benchmark. Built using our agentic flywheel, it has been trained to execute real-world workflows within synthetic enterprise environments. This not only ensures that Holo3 excels in today's business scenarios, but establishes the foundation for a future where our agents can autonomously navigate virtually any digital landscape. Best of all, Holo3 achieves this with only 10B active parameters (122B total), so at a fraction of the cost of large-scale proprietary models, such as GPT 5.4 or Opus 4.6.
The Download: gig workers training humanoids, and better AI benchmarks
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. The gig workers who are training humanoid robots at home When Zeus, a medical student in Nigeria, returns to his apartment from a long day at the hospital, he straps his iPhone to his forehead and records himself doing chores. Zeus is a data recorder for Micro1, which sells the data he collects to robotics firms. As these companies race to build humanoids, videos from workers like Zeus have become the hottest new way to train them. Micro1 has hired thousands of them in more than 50 countries, including India, Nigeria, and Argentina.
Falcon Perception
On SA-Co, Falcon Perception reaches 68.0 Macro-F1 (vs. 62.3 for SAM 3) with the main remaining gap being presence calibration (MCC 0.64 vs. 0.82). We also relase Falcon OCR, a 0.3B-parameter model which reaches a score of 80.3 and 88.6 on the olmOCR benchmark and OmniDocBench respectively, while having the highest throughput of any open source OCR model. Many open-vocabulary perception systems are built as modular pipelines: a (often frozen) vision backbone extracts features, a separate fusion/decoder stage combines them with language, and additional components handle matching and post-processing. One Backbone, Two Behaviors At its core, Falcon Perception is a dense Transformer that processes image patches and text tokens in a shared parameter space from the first layer.
Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents
Granite 4.0 3B Vision excels on the following capabilities: The model ships as a LoRA adapter on top of Granite 4.0 Micro, our dense language model, keeping vision and language modular for text-only fallbacks and seamless integration into mixed pipelines. How Granite 4.0 3B Vision Was Built Granite 4.0 3B Vision’s performance is the result of three key investments: A purpose-built chart understanding dataset constructed via a novel code-guided data augmentation approach, a novel variant of the DeepStack architecture that enables high-detail visual feature injection, and a modular design that keeps the model practical for enterprise deployment. To close this gap, we’ve developed ChartNet: a million-scale multimodal dataset purpose-built for chart interpretation and reasoning, described in detail in our upcoming CVPR 2026 paper.
Shifting to AI model customization is an architectural imperative
In the early days of large language models (LLMs), we grew accustomed to massive 10x jumps in reasoning and coding capability with every new model iteration. When a model is fused with an organization’s proprietary data and internal logic, it encodes the company’s history into its future workflows. This alignment creates a compounding advantage: a competitive moat built on a model that understands the business intimately. Integrated into Mistral’s software development scaffolding, this customized model now supports the entire lifecycle—from maintaining legacy systems to autonomous code modernization via reinforcement learning. Reliance on a single cloud provider or vendor for model alignment creates a dangerous asymmetry of power regarding data residency, pricing, and architectural updates.
AI benchmarks are broken. Here’s what we need instead.
From chess to advanced math, from coding to essay writing, the performance of AI models and applications is tested against that of individual humans completing tasks. This framing is seductive: An AI vs. human comparison on isolated problems with clear right or wrong answers is easy to standardize, compare, and optimize. Although researchers and industry have started to improve benchmarking by moving beyond static tests to more dynamic evaluation methods, these innovations resolve only part of the issue. This misalignment leaves us misunderstanding AI’s capabilities, overlooking systemic risks, and misjudging its economic and social consequences. I have studied real-world AI deployment since 2022 in small businesses and health, humanitarian, nonprofit, and higher-education organizations in the UK, the United States, and Asia, as well as within leading AI design ecosystems in London and Silicon Valley.
TRL v1.0: Post-Training Library That Holds When the Field Invalidates Its Own Assumptions
This is what we tried to solve in TRL v1.0, and this post explains how. 1. A moving target: post-training as a shifting field Post-training has not evolved as a smooth refinement of one recipe. PPO [Schulman et al., (2017); Ziegler et al., (2019)] made one architecture look canonical: a policy, a reference model, a learned reward model, sampled rollouts, and an RL loop. Then DPO-style methods such as the original DPO [Rafailov et al., (2023)], ORPO [Hong et al., (2024)], and KTO [Ethayarajh et al., (2024)] cut through that stack: preference optimization could work without a separate reward model, value model, or any online RL.
Secure governance accelerates financial AI revenue growth
Financial institutions are learning to deploy compliant AI solutions for greater revenue growth and market advantage. During that era, quantitative teams programmed systems designed to discover ledger discrepancies or eliminate milliseconds from automated trading execution times. Today, it’s not acceptable for banking executives to approve new technology rollouts based simply on promises of accurate predictive capabilities. Consequently, the dialogue within corporate boardrooms has narrowed intensely to focus on safe AI deployment, ethics, model oversight, and legislation specific to the financial industry. Mastering these requirements creates a highly efficient operational pipeline where good governance functions as a massive accelerant for product delivery rather than an administrative handbrake. Consider a scenario where a multinational bank introduces a deep learning framework to process commercial loan applications.
A woman’s uterus has been kept alive outside the body for the first time
“Think of this as a human body,” says Javier González. Ten months ago, González, a biomedical scientist who developed the device with his colleagues at the Carlos Simon Foundation, carefully placed a freshly donated human uterus in the tub. The device kept the uterus alive for a day—a new feat that could represent the first step to the long-term maintenance of uteruses outside the human body. The work has not yet been published. The team members want to keep donated human uteruses alive long enough to see a full menstrual cycle. Mother’s first uterus The team first began testing an early prototype of the device with sheep uteruses around four years ago.
Here’s why some people choose cryonics to store their bodies and brains after death
This week I reported on some rather unusual research that focuses on the brain of L. Coles was a gerontologist who died from pancreatic cancer in 2014. And before he died, he decided to have his brain preserved by a cryonics facility. Today, it’s being stored at −146 °C at a center in Arizona, where it sits covered in a thin layer of frost. Over the past few years, I’ve spoken to people who run cryonics facilities, study cryopreservation, or just want to be cryogenically stored. Alcor charges $80,000 to store a person’s brain, and around $220,000 to store a whole body.
Watch James Manyika talk AI and creativity with LL COOL J.
In the latest episode of our Dialogues on Technology and Society series, legendary artist LL COOL J sits down with James Manyika, Google's Senior Vice President of Research, Labs, Technology & Society. They have a wide-ranging conversation on the evolution of creativity and technology. LL reflects on his 40-year career, witnessing technology evolve from the first drum machines to generative AI. He discusses the potential of AI to democratize access for a new generation of artists, while emphasizing the importance of protecting the “divine spark” that makes creativity human.
Training Driving AI at 50,000× Real Time
An automated system must interpret a chaotic, ever-changing world in real time—navigating uncertainty, predicting human behavior, and operating safely across an immense range of environments and edge cases. At General Motors, we approach this problem from a simple premise: while most moments on the road are predictable, the rare, ambiguous, and unexpected events — the long tail — are what ultimately defines whether an autonomous system is safe, reliable, and ready for deployment at scale. GM is building scalable driving AI to meet that challenge — combining large-scale simulation, reinforcement learning, and foundation-model-based reasoning to train autonomous systems at a scale and speed that would be impossible in the real world alone. Stress-testing for the long tail Long-tail scenarios of autonomous driving come in a few varieties.
Roundtables: The Next Era of Space Exploration
Listen to the session or watch below Whether it's the race to find life on Mars, the campaign to outsmart killer asteroids, or the quest to make the moon a permanent home to astronauts, scientists' efforts in space can tell us more about where humanity is headed. Speakers: Amanda Silverman, editor, features & Investigations, and Robin George Andrews, award-winning science journalist and author Recorded on March 25, 2026 Related Stories: Backlash against ICE is fueling a broader movement against AI companies’ ties to President Trump. Exclusive: Niantic's AI spinout is training a new world model using 30 billion images of urban landmarks crowdsourced from players.
Exclusive eBook: Are we ready to hand AI agents the keys?
We’re starting to give AI agents real autonomy, but are we prepared for what could happen next? This subscriber-only eBook explores this and angles from experts, such as “If we continue on the current path … we are basically playing Russian roulette with humanity.” by Grace Huckins June 12, 2025 Related Stories: Access all subscriber-only eBooks: Backlash against ICE is fueling a broader movement against AI companies’ ties to President Trump. Exclusive: Niantic's AI spinout is training a new world model using 30 billion images of urban landmarks crowdsourced from players. They argue we need a revolution—and more and more influential scientists, funders, and politicians are taking them seriously.
This scientist rewarmed and studied pieces of his friend’s cryopreserved brain
Stephen Coles’s brain sits cushioned in a vat at a storage facility in Arizona. Before he died, he asked cryobiologist Greg Fahy to study the effects of the preservation procedure on his brain. Coles was especially curious about whether his cooled brain would crack, says Fahy. Coles’s brain was preserved shortly after he died in 2014, but Fahy has only recently got around to analyzing those samples. He says that Coles’s brain is “astonishingly well preserved.” “We can see every detail [in the structure of the brain biopsies],” says Fahy, who is chief scientific officer at biotech companies Intervene Immune and 21st Century Medicine (where he is also executive director).
Securing AI systems under today’s and tomorrow’s conditions
However, there are security risks to building models and training them on that data. The eBook’s authors state that organisations need to manage threats throughout their AI development and implementation processes. Utimaco lists three areas under threat: Current public key cryptography will become vulnerable in the next ten years, the report’s authors attest; a period in which capable quantum systems may emerge. The report’s authors suggest what they term ‘crypto-agility’, which it defines as changing cryptographic algorithms without redesigning underlying systems. ‘Crypto-agility’ is based on the principle of hybrid cryptography – combining established algorithms with post-quantum methods, such as those suggested by NIST.
The Download: tracing AI-fueled delusions, and OpenAI admits Microsoft risks
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. To find out, Stanford researchers analyzed transcripts from chatbot users who experienced these spirals. Their findings suggest that chatbots have a unique ability to turn a benign, delusion-like thought into a dangerous obsession. (CNBC) + OpenAI is wooing private equity firms with a sweeter deal than Anthropic’s. (Reuters $) + It’s also building a fully automated researcher. (MIT Technology Review) + And wants to muscle in on Google’s search dominance. (Telegraph $) 2 The US just banned all new foreign-made consumer routers Citing national security concerns.
Why Thermal Metrology Must Evolve for Next-Generation Semiconductors
This whitepaper provides engineers and researchers with a technical framework for understanding how thermal behavior is changing in advanced semiconductor systems. As devices scale into three-dimensional architectures with thinner layers, higher power densities, and increasingly complex material stacks, heat transport becomes confined, interface-dominated, and highly sensitive to small variations in structure and processing. These shifts are making accurate thermal characterization increasingly important for design workflows, model validation, and long-term reliability in modern electronic systems. What you will learn about: Why classical bulk thermal assumptions break down at nanometer-scale film thicknesses and how this affects thermal conductivity and device-level modeling accuracy.
The Bay Area’s animal welfare movement wants to recruit AI
In early February, animal welfare advocates and AI researchers gathered in stocking feet at Mox, a scrappy, shoes-free coworking space in San Francisco. In front of the “Bovine Room” stood a bookshelf stacked with copies of Eliezer Yudkowsky’s If Anyone Builds It, Everyone Dies, a manifesto arguing that AI could wipe out humanity. The event was hosted by Sentient Futures, an organization that believes the future of animal welfare will depend on AI. Like many Bay Area denizens, the attendees were decidedly “AGI-pilled”—they believe that artificial general intelligence, powerful AI that can compete with humans on most cognitive tasks, is on the horizon. One person pitched a $100 million animal super PAC that would place staffers with Congress members and lobby for animal welfare legislation.
What Happens If AI Makes Things Too Easy for Us?
But researchers are concerned AI is making some tasks too easy, and that this will come with unexpected costs. In a commentary titled Against Frictionless AI, published in Communications Psychology on 24 February, psychologists from the University of Toronto discuss what might be lost when AI removes too much effort from human activities. Their argument centers on the idea that friction—difficulty, struggle, and even discomfort—plays an important role in learning, motivation, and meaning. In the context of work, it involves mental effort—rumination and persistence, staying on a problem for some time, and this helps solidify the idea and the creative process.
Build a Domain-Specific Embedding Model in Under a Day
Back to Articles Build a Domain-Specific Embedding Model in Under a Day Enterprise + Article Published March 20, 2026 Upvote 2 Steve H steve-nvidia Follow nvidia Rucha Apte ruchaa01 Follow nvidia Sean Sodha ssodha-nv Follow nvidia Oliver Holworthy nvidia-oliver-holworthy Follow nvidia If you are building a RAG (Retrieval-Augmented Generation) system, you have likely hit this wall: Everything works… until it doesn’t. Fine-tuning an embedding model can improve the performance of your retrieval pipeline when off-the-shelf models fail to effectively capture domain-specific nuances. With a single GPU and less than a day of training time, you can transform a general-purpose embedding model into one that truly understands your domain, no manual labeling required. To get started, follow this setup guide. 📚 Step 1: Generate Training Data from Documents Fine-tuning an embedding model requires thousands of (query, relevant document) pairs.